A personalized path recommendation method based on the A* algorithm guided by deep learning

Abstract

The present invention belongs to the field of personalized path recommendation algorithms in the direction of deep learning, and relates to a personalized path recommendation method based on the A* algorithm guided by deep learning, including: preprocessing data and extracting available data from the user's historical travel information; the CNN+LSTM neural network learns the implicit personalized preference features of the user through the extracted data, uses the CNN+LSTM neural network for path planning to obtain alternative paths; selects the evaluation criteria of the path as the estimation function of the A* search algorithm, which is used as the basis for selecting the next moment node in the iterative process of the A* search algorithm; the path planned by the A* search algorithm is the final solution result, that is, the personalized path recommended for the user. This algorithm provides a new option for the route recommendation function of modern navigation systems, can provide route suggestions that better meet the needs of users, improve the satisfaction of users with navigation services, and improve the quality of traffic travel.

Description

Technical Field

[0001] The present invention belongs to the field of personalized path recommendation algorithms in the direction of deep learning, and specifically relates to a personalized path recommendation method based on the A* algorithm guided by deep learning. Background Art

[0002] Transportation is an essential part of people's daily lives. Travel methods and accessibility are directly related to improving people's quality of life. With the development of the times and advancements in technology, transportation methods have undergone tremendous changes. Traditional transportation methods have been replaced by modern ones. In the past, people wasted precious time on the road, and travel plans were often disrupted by various emergencies due to long journeys, causing inconvenience to people's lives. Since the 21st century, rapid economic development has driven rapid development in the transportation sector, significantly increasing the level of information technology in travel. Comprehensive and comprehensive transportation information services can provide people with information on traffic conditions, transportation service facilities, traffic control, weather conditions, public transportation services, and tourism and entertainment services, greatly improving the comfort, convenience, and safety of people's travel and meeting the personalized needs of the public.

[0003] Current transportation services are far from perfect, with significant room for improvement in many areas. Route recommendation, a crucial component of transportation services, is also a core function of many current navigation systems, such as Baidu Maps and Amap. Route recommendation services can provide users with targeted travel suggestions that meet their specific needs, significantly optimizing their travel experience. Research on route recommendation algorithms and related fields has become a key focus for researchers in the field of map-based routing applications.

[0004] As society continues to develop, people's material lives are becoming vastly more enriched and their living standards are constantly improving. This has led to increasingly complex and diverse demands, creating a more pressing need for improved service quality. People are increasingly demanding the personalization, timeliness, and efficiency of traffic information, placing new demands on the quality of route recommendation services. The continuous advancement of urban construction, the increasing complexity of road networks, and the ever-changing traffic conditions further increase the challenges of route recommendation services.

[0005] Most early studies regarded the path recommendation task as a path - finding problem on a graph. These studies mainly focused on how to extend existing algorithms, such as Dijkstra's shortest path algorithm and A* search algorithm. Most of these methods are based on static path - planning methods, that is, according to the user's travel trajectory in a certain historical period in the past, a model is established for the user and based on this, the routes that the user may choose in the future are planned. In actual application scenarios, user behavior is often affected by multiple factors. The user's choice preferences change with the change of influencing factors. These methods do not take into account the timeliness of user behavior and cannot well cope with the dynamic changes of user needs and preferences. Therefore, analyzing the relationship between user choices, user needs, and traffic conditions based on the user's historical behavior and road network characteristics and constructing a new path - recommendation algorithm have important practical significance for solving the dynamic personalized path - recommendation task. Summary of the Invention

[0006] Aiming at the main problems existing in the personalized path - recommendation algorithms currently applied in the field of transportation, the present invention combines deep - learning technology and the A* algorithm to propose an improved personalized path - recommendation method - the CNN - LSTM - A* method. This personalized path - recommendation algorithm provides a new option for the route - recommendation function of modern navigation systems, can provide route suggestions that better meet the user's needs, improve the user's satisfaction with navigation services, and improve the quality of transportation.

[0007] For the task of personalized path recommendation, it is specifically described as follows:

[0008] There are a certain number and size of regions in the road network G. According to the influence of these regions on user behavior, the regions in G can be divided into the following two categories: obstacles and non - obstacles. For obstacles, users cannot enter their range; for non - obstacles, users can enter their range. Considering the user's personalized characteristics, non - obstacles can be further divided into regions that users tend to enter and regions that users tend to avoid. For regions that users tend to enter, when choosing a path, users will try to pass through these regions as much as possible, that is, users will try to extend the path length in these regions; for regions that users tend to avoid, when choosing a path, users will try to avoid these regions as much as possible, that is, users will try to shorten the path length in these regions. The solution result of the algorithm should avoid obstacles and conform to the preference characteristics of users for different types of regions in the historical behavior records.

[0009] The technical solution of the present invention is as follows:

[0010] A personalized path - recommendation method based on a deep - learning - guided A* algorithm, abbreviated as the CNN - LSTM - A* method, mainly consists of two components: the CNN + LSTM neural network and the A* search algorithm; specifically as follows:

[0011] (1) First, preprocess the data and extract usable data from the user's historical travel information;

[0012] The CNN-LSTM-A* method requires data preprocessing of historical user behavior information and the new environment for route recommendations. In practical applications, road network and route data can take many forms. The algorithm uses image feature extraction techniques based on the specific form of the data provided to simplify it into a unified two-dimensional image. The algorithm processes the image by geometrically graphing the object along its image boundaries.

[0013] (2) The CNN+LSTM neural network learns the implicit personalized preference features of users from the extracted data, and uses the CNN+LSTM neural network to perform path planning and obtain alternative paths;

[0014] The CNN+LSTM neural network component of the CNN-LSTM-A* method learns and utilizes personalized features. First, the CNN+LSTM neural network component is trained and tested to learn personalized features. Based on actual application needs, the CNN-LSTM-A* method divides pre-processed user historical data into training and test sets. The CNN+LSTM neural network component learns personalized features by processing the training set data. The positional relationship between the upper node and the non-obstacle object is input as training data into the CNN+LSTM neural network component, which then extracts features from this positional relationship. After training, the CNN+LSTM neural network component generates a set of neural network parameters, which serve as the output data of the CNN+LSTM neural network component. The CNN-LSTM-A* method uses the test set to evaluate the performance of the CNN+LSTM neural network component. When calling the CNN-LSTM-A* method, you can use the evaluation results of the CNN+LSTM neural network component to debug the neural network model.

[0015] The CNN-LSTM-A* method performs path planning based on the CNN+LSTM neural network component. T In the current state s of the current node agent at time t t The positional relationship between the agent and the non-obstacle is input as input data into the CNN+LSTM neural network component, which will predict the agent's behavior a at the current time t. t , and then according to a t Determine the state s of the agent at the next moment t+1 , thus obtaining the planned path.

[0016] (3) The evaluation criteria for alternative paths serve as the estimation function of the A* search algorithm and are used as the basis for selecting the next moment node in the iterative process of the A* search algorithm; the path planned by the A* search algorithm is the final solution of the CNN-LSTM-A* method, that is, the personalized path recommended for the user.

[0017] The CNN+LSTM neural network component can solve to obtain a complete reachable path. This path only considers meeting personalized features and cannot guarantee global optimality and local optimality. The CNN-LSTM-A* method evaluates the path planning results of the CNN+LSTM neural network, and the values of the evaluation metrics will be applied to the A* search algorithm. The CNN-LSTM-A* method applies the trained neural network component in the A* search algorithm to implement the use of personalized features. The A* search algorithm conducts path search based on the evaluation value of the solution result of the CNN+LSTM neural network component and solves to obtain a reachable path within the global scope, that is, the solution to the personalized path recommendation problem.

[0018] The specific steps of the present invention are as follows:

[0019] Step S1: Data preprocessing. Unify the form and collect data for the original input data.

[0020] Step S2: The CNN+LSTM neural network component learns the user's historical behavior data.

[0021] Step S3: Add the end point B T to the open list of the A* search algorithm, and take agent = B T .

[0022] Step S4: Select the node with the smallest function value of the evaluation function f(n) in the open list and move it from the open list to the close list.

[0023] Step S5: Update the agent.

[0024] Step S6: Expand the adjacent nodes of the agent.

[0025] Step S7: Add the reachable adjacent nodes that are not in the close list to the open list.

[0026] Step S8: Handle the following three situations separately: If step S7 adds E T to the open list, the personalized path solution is successful and the algorithm terminates; if the open list is empty, the personalized path solution fails and the algorithm terminates; if neither of the above two situations is satisfied, repeat steps S4 to S7 until the algorithm termination condition is reached.

[0027] The beneficial effects of the present invention:

[0028] (1) Belongs to the personalized route recommendation algorithm and is applied to the personalized path recommendation task in the field of transportation.

[0029] (2) Can better cope with the influencing factors of the recommendation results, the dynamic changes of users' personalized preferences and user needs.

[0030] (3) Can be applied to application scenarios with complex road network conditions and can consider multiple factors that can affect the recommendation results at the same time. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] Figure 1 Is a flowchart of image processing.

[0032] Figure 2 Is a structure diagram of the CNN+LSTM neural network.

[0033] Figure 3 Is an execution flowchart of the A* search algorithm.

[0034] Figure 4 Is a schematic diagram of users' historical travel information.

[0035] Figure 5 Is the path planning result of the CNN navigator in the circular obstacle environment.

[0036] Figure 6 Is the path planning result of the CNN navigator in the square obstacle environment.

[0037] Figure 7 Is the path planning result of the CNN-LATM navigator in the circular obstacle environment.

[0038] Figure 8 Is the path planning result of the CNN-LSTM navigator in the square obstacle environment.

[0039] Figure 9 Is the path recommendation result of the CNN-A* model in the circular obstacle environment.

[0040] Figure 10 Is the path recommendation result of the CNN-LSTM-A* model in the circular obstacle environment. DETAILED DESCRIPTION OF THE INVENTION

[0041] The following further describes the specific implementation manners of the present invention in conjunction with the accompanying drawings.

[0042] The process of processing images in the data preprocessing part of the present invention is as Figure 1 shown, and the specific steps are described as follows:

[0043] Step 1: Through image feature extraction technology, the algorithm extracts the edges of the research object in the original image and forms geometric figures to determine the road network G to be recommended. T , the scope of the historical road network G0 and G T , the range of obstacles and non-obstacles in G0, and determine the user's historical travel path The trajectory curve.

[0044] Step 2: Establish a two-dimensional coordinate system, define nodes, and determine G T , the position coordinates of each node within the G0 range.

[0045] Step 3: According to the actual meaning of the research object in the image, G T , add corresponding category labels to the geometric figures within the G0 range.

[0046] Step 4: Collection The position relationship between each node and non-obstacles.

[0047] The structure of CNN+LSTM neural network is as follows Figure 2 The neural network component of the CNN-LSTM-A* method is a combination of a convolutional neural network (CNN) and a long short-term memory (LSTM). The LSTM is inserted after the pooling layer and before the fully connected layer of the CNN. The CNN+LSTM neural network component can be specifically divided into the following four layers:

[0048] First, the CNN+LSTM neural network component uses a CNN to encapsulate a set of convolutional layers, activation layers, and pooling layers into a single CNN layer. The CNN-LSTM neural network component sets up three CNN layers, using the leakyRelu activation function. Then, the CNN+LSTM neural network component uses the LSTM to encapsulate multiple LSTM layers into a single LSTM layer. The CNN+LSTM neural network component sets up an LSTM layer after the encapsulated CNN layer. The LSTM layer contains a recurrent neural network layer. Next, the CNN+LSTM neural network component sets up another activation layer, using the leakyRelu activation function. Finally, the CNN+LSTM neural network component sets up the fully connected layers for the remaining CNN layers. The CNN+LSTM neural network component sets up two fully connected layers.

[0049] The execution process of the A* search algorithm is as follows Figure 3 As shown. The iteration of the A* search algorithm from time t to time t+1 can be divided into the following three steps:

[0050] First, for the agent at time t, expand its reachable adjacent nodes within its neighborhood that are not in the close list and add them to the open list. Here, reachable means that neither the node nor any point on the straight line connecting the node and the agent is within the obstacle range. The A* search algorithm iterates based on the open list and the close list. For an open list with k nodes and a close list with m nodes, their specific expressions are shown in Formulas (1) and (2):

[0051] open list = {n1, n2, …, n k} (1)

[0052] close list = {n1, n2, …, n m} (2)

[0053] The specific iteration formula for the first step is shown in Formula (3):

[0054] f(n ξ ) = min(f(n1), f(n2), …, f(n k ))), ξ ∈ [1, k], n1, n2, …, n k ∈ open list (3)

[0055] Then the A* search algorithm selects a node n ξ from the open list and adds it to the close list. The node n ξ needs to meet the following conditions: the function value of f(n ξ ) is the smallest, it is reachable, and the node n ξ is not in the close list.

[0056] For the A* search algorithm, g(n) is the movement cost from the starting point to the current node, h(n) is the estimated cost from the current node to the end point, and the calculation formula of the evaluation function f(n) is shown in Formula (4):

[0057] f(n) = g(n) + h(n) (4)

[0058] Then at time t, for the agent, its adjacent nodes n1, n2, …, n k ∈ N t . g(agent) is the evaluation value of the partial path that has been solved from B T to the agent at time t, h(agent) is the evaluation value of the path planned by the neural network component from the agent to E T at time t, and the calculation formula of the evaluation function f(agent) of the agent is shown in Formula (5):

[0059] f(agent)=g(agent)+h(agent) (5)

[0060] Then at time t, for an adjacent node n of the agent ξ , n ξ ∈N t ,ξ∈[1,k]. Define D ξ Indicates that at time t, the agent is the starting point and n ξ is the evaluation value of the straight line path at the end point, g(n ξ ) is calculated as shown in formula (6):

[0061] g(n ξ )=g(agent)+D ξ (6)

[0062] h(n ξ ) is the time from n ξ to E T The evaluation value of the path planned by the neural network component, n ξ The calculation formula of the evaluation function is shown in formula (7):

[0063] f(n ξ )=g(n ξ )+h(n ξ ) (7)

[0064] The specific iterative formulas for the second step are shown in formulas (8) and (9):

[0065] open list = open list - {n ξ} (8)

[0066] close list = close list + {n ξ} (9)

[0067] Finally update the agent, the selected node n ξ is the agent at time t+1.

[0068] The specific iterative formula of the third step is shown in formula (10):

[0069] agent=n ξ (10)

[0070] The user's historical travel information simulated in the experiment is as follows Figure 4 As shown in the figure, the CNN-LSTM-A* method requires data preprocessing of user historical behavior information and extracting usable data from user historical travel information. Historical behavior information includes the historical road network environment G0 and the user's historical path The polyline represents the user's historical travel trajectory, the dark circles represent obstacles, the light circles represent objects, the solid dot with coordinates (0.0, 0.0) represents the starting point, the solid dot with coordinates (1.0, 1.0) represents the ending point, and the solid dots on the polyline represent the nodes on the path. The obstacle labels are known information provided by the input data. By bypassing all the light circles along the trajectory, it can be seen that the objects in the historical travel data are areas that the user tends to avoid.

[0071] The experiments respectively use the CNN navigator and the CNN-LSTM navigator to perform path planning in obstacle environments with different shapes. The visualization results are as Figures 5 to 8 shown. The polyline represents the path recommended by the model, the dark circles or squares represent obstacles, the light circles or squares represent objects, the solid dot with coordinates (0.0, 0.0) represents the starting point, the solid dot with coordinates (1.0, 1.0) represents the ending point, and the solid dots on the polyline represent the nodes on the path. The obstacle labels are known information provided by the input data, and the labels added by the navigator to the objects are the output results of the trained neural network components.

[0072] The path recommendation results of the CNN-A* model and the CNN-LSTM-A* model are as Figure 9 and Figure 10 shown. The experiments establish the CNN-A* model based on the CNN-A* algorithm and the CNN-LSTM-A* model based on the CNN-LSTM-A* method. Based on the user's historical travel information as Figure 4 shown, the experiments test the performance of the CNN-LSTM-A* method by comparing the path recommendation results of the above models. The paths recommended by the CNN-A* model and the CNN-LSTM-A* model in the figure are concentrated and circuitous near the optimal path. Compared with the path recommended by the CNN-A* model, the path recommended by the CNN-LSTM-A* model has a smaller average distance to the optimal path and is more evenly distributed globally. The above visualization results show that compared with the CNN-A* method, the path recommendation result of the CNN-LSTM-A* method has a shorter path length and a smoother path on the premise of meeting the user's personalized needs, and the CNN-LSTM-A* method is more excellent in the personalized path recommendation task.

Claims

1. A personalized path recommendation method based on the A* algorithm guided by deep learning, called the CNN-LSTM-A* method, is characterized in that It is mainly composed of two components: the CNN+LSTM neural network and the A* search algorithm; specifically as follows: (1) First, preprocess the data and extract available data from the user's historical travel information; According to the specific form of the provided data, use the corresponding image feature extraction technology to uniformly simplify the data into image information in a two-dimensional coordinate system; (2) The CNN+LSTM neural network learns the implicit personalized preference features of the user from the extracted data, and uses the CNN+LSTM neural network for path planning to obtain alternative paths; First, train and test the CNN+LSTM neural network component to achieve learning of personalized features; According to the needs of actual applications, the preprocessed user historical data is reasonably divided into a training set and a test set; the CNN+LSTM neural network component learns user personalized features by processing the training set data; the user historical path The positional relationship between the upper node and the non-obstacle is used as the training set data to input into the CNN+LSTM neural network component, and the CNN+LSTM neural network component extracts the features of its positional relationship; The CNN+LSTM neural network component obtains a set of neural network parameters after training as the output data of the CNN+LSTM neural network component; the CNN-LSTM-A* method uses the test set to evaluate the performance of the CNN+LSTM neural network component. When calling the CNN-LSTM-A* method, debug the neural network model according to the evaluation results of the CNN+LSTM neural network component; In the road network G of the new environment T the state s of the current node agent at the current moment t t and the positional relationship between the agent and non - obstacles are input as input data into the CNN + LSTM neural network component. The CNN + LSTM neural network component will predict the behavior a of the agent at the current moment t t and then, based on a t determine the state s of the agent at the next moment t+1 so as to obtain the planned path; (3) The evaluation criteria for alternative paths are used as the estimation function of the A* search algorithm, which is used as the basis for selecting the next moment node in the iterative process of the A* search algorithm; the path planned by the A* search algorithm is the final solution of the CNN-LSTM-A* method, that is, the personalized path recommended for the user; Evaluate the path planning results of the applied CNN+LSTM neural network, and the values of the evaluation indicators are applied to the A* search algorithm; the A* search algorithm performs path search based on the evaluation values of the solution results of the CNN+LSTM neural network component, and solves to obtain reachable paths within the global range, that is, the solution to the personalized path recommendation problem.

2. The personalized path recommendation method based on the A* algorithm guided by deep learning according to claim 1, wherein, In the described CNN+LSTM neural network, LSTM is inserted after the pooling layer and before the fully connected layer of CNN; specifically divided into the following four layers: First, the CNN+LSTM neural network component uses CNN to encapsulate a set of convolutional layers, activation function layers, and pooling layers of CNN into one CNN layer. The CNN-LSTM neural network component sets three CNN layers, and the activation function is the leakyRelu function; then, the CNN+LSTM neural network component uses LSTM to encapsulate multiple layers of LSTM into one LSTM layer; the CNN+LSTM neural network component sets the LSTM layer after the encapsulated CNN layer, and the LSTM layer contains one layer of recurrent neural network; then, the CNN+LSTM neural network component sets the activation function layer again, and the activation function is the leakyRelu function; finally, the CNN+LSTM neural network component sets the fully connected layers of the remaining part of CNN, and sets two fully connected layers.

3. The personalized path recommendation method based on the A* algorithm guided by deep learning according to claim 1 or 2, characterized in that, The described A* search algorithm is divided into the following three steps for the iteration from time t to time t+1: First, for the agent at time t, expand its reachable adjacent nodes within its neighborhood that are not in the close list and add them to the open list. Here, reachable means that neither the node nor any point on the straight line connecting the node and the agent is within the obstacle range. The A* search algorithm iterates based on the open list and the close list. For an open list with k nodes and a close list with m nodes, their specific expressions are shown in Formulas (1) and (2): open list = {n1, n2, …, n k} (1) close list = {n1, n2, …, n m} (2) The specific iteration formula for the first step is shown in Formula (3): f(n ξ ) = min(f(n1), f(n2), …, f(n k )),ξ ∈ [1, k],n1, n2, …, n k ∈ open list (3) Then the A* search algorithm selects a node n from the open list ξ and adds it to the closed list; the node n ξ must satisfy the following conditions: the function value of f(n ξ ) is the smallest, it is reachable, and the node n ξ is not in the closed list; For the A* search algorithm, g(n) is the movement cost from the starting point to the current node, h(n) is the estimated cost from the current node to the end point, and the calculation formula for the evaluation function f(n) is shown in Formula (4): f(n) = g(n) + h(n) (4) At time t, for the agent, its adjacent nodes n1, n2, …, n k ∈ N t ; g(agent) is the evaluation value of the partially solved path from B T to the agent at time t, h(agent) is the evaluation value of the path planned by the neural network component from the agent to E T at time t, and the calculation formula of the evaluation function f(agent) of the agent is shown in formula (5): f(agent) = g(agent) + h(agent) (5) At time t, for an adjacent node n of the agent ξ , n ξ ∈N t , ξ ∈ [1, k]; Define D ξ to represent the evaluation value of the straight-line path starting from the agent at time t and ending at n ξ . The calculation formula of g(n ξ ) is shown in formula (6): g(n ξ ) = g(agent) + D ξ (6) h(n ξ ) is the evaluation value of the path planned by the neural network component from n ξ to E T at time t. The calculation formula of the evaluation function of n ξ is shown in Formula (7): f(n ξ ) = g(n ξ ) + h(n ξ ) (7) The specific iteration formulas for the second step are shown in Formulas (8) and (9): open list = open list - {n ξ} (8) close list = close list + {n ξ} (9) Update the agent last, the selected node n ξ is the agent at time t + 1; The specific iteration formula for the third step is shown in Formula (10): agent=n ξ (10).

4. The personalized path recommendation method based on the A* algorithm guided by deep learning according to claim 3, characterized in that, The specific steps of the CNN-LSTM-A* method are as follows: Step S1: Data preprocessing, unify the form and collect data for the original input data; Step S2: The CNN+LSTM neural network component learns the user's historical behavior data; Step S3: Add the end point B T to the open list of the A* search algorithm, and set agent = B T ; Step S4: Select the node with the smallest function value of the evaluation function f(n) in the open list, and move it from the open list to the close list; Step S5: Update the agent; Step S6: Expand the adjacent nodes of the agent; Step S7: Add the reachable adjacent nodes that are not in the close list to the open list; Step S8: Handle the following three cases separately: If step S7 adds E T to the open list, the personalized path solving is successful and the algorithm terminates; if the open list is empty, the personalized path solving fails and the algorithm terminates; if neither of the above two cases is satisfied, repeat steps S4 - S7 until the algorithm termination condition is reached.

Images